L01 Mini PDF

Introduction to Optimal Control via LMIs
Matthew M. Peet
Arizona State University
Lecture 01: Optimal Control via LMIs

Summary
Powerful New Tools

• Convex Optimization
I LMIs
I Sum-of-Squares
Many old problems have been solved

• H∞ and H2 optimal control
• Nonlinear stability analysis
• Analysis and Control of delayed and PDE systems
Many questions are still unresolved

• Control of nonlinear Systems
• Nonlinear Programming (partially resolved)
Question: What is meant by a “solution”?
M. Peet Lecture 01: 1 / 135

Outline
Lectures 1-2
1. Linear Systems
2. Convex Optimization and Linear Matrix Inequalities
3. Optimal Control
4. LMI Solutions to the H∞ and H2 Optimal Control Problems

Signal Spaces
L2 space
Definition 1.
L2 [0, ∞) is the Hilbert space of functions f : R+ → Rn with inner product
Z ∞
1
hu, yiL2 = u(t)T v(t)dt
2π 0
L2 [0, ∞) inherits the norm

Z ∞
kuk2L2 = ku(t)k2 dt
0

Operator Theory
Linear Operators
Definition 2.
The normed space of bounded linear operators from X to Y is denoted
L(X, Y) with norm
kP xkY
kP kL(X,Y ) := sup =K
x∈X kxkX
x6=0
• Satisfies the properties of a norm

• This type of norm is called an “induced” norm
• Notation: L(X) := L(X, X)
• If X is a Banach space, then L(X, Y ) is a Banach space
Properties: Suppose G1 ∈ L(X, Y ) and G2 ∈ L(Y, Z)
• Then G2 G1 ∈ L(X, Z).
• kG2 G1 kL(X,Z) ≤ kG2 kL(Y,Z) kG1 kL(X,Y ) .
• Composition forms an algebra.
Laplace Transform
Definition 3.
Given u ∈ L2 [0, ∞), the Laplace Transform of u is û = Λu, where
Z T
û(s) = (Λu)(s) = lim u(t)e−st dt
T →∞ 0
if this limit exists.

Λ is a bounded linear operator - Λ ∈ L(L2 , H2 ).
• Λ : L2 → H2 .
• The norm kΛkL(L2 ,H2 ) is
kΛukH2
kΛk = sup =???
u∈L2 kukL2

H2 - A Space of Integrable Analytic Functions
Definition 4.
A complex function is analytic if it is continuous and bounded.
A function is analytic if the Taylor series converges everywhere in the domain.
Definition 5.
A function û : C̄+ → Cn is in H2 if
1. û(s) is analytic on the Open RHP (denoted C+ )
2. For almost every real ω,
lim û(σ + ıω) = û(ıω)

σ→0+
I Which means continuous on the imaginary axis

3. Z ∞
supkû(σ + ıω)k22 < ∞
−∞ σ≥0
I Which means integrable on every vertical line.

The Maximum Modulus Principle
Theorem 6 (Maximum Modulus).

An analytic function cannot obtain its extrema in the interior of the domain.
Hence if û satisfies 1) and 2), then

Z ∞ Z ∞
2
supkû(σ + ıω)k2 = kû(ıω)k22 dω
−∞ σ≥0 −∞
We equip H2 with a norm and inner product

Z ∞ Z ∞
1
kûkH2 = 2
kû(ıω)k2 dω, hû, ŷiH2 = û(ıω)∗ v̂(ıω)dω
−∞ 2π −∞

Paley-Wiener
Theorem 7.
1. If u ∈ L2 [0, ∞), then Λu ∈ H2 .
2. If û ∈ H2 , then there exists a u ∈ L2 [0, ∞) such that û = Λu ( Onto).
• Shows that H2 is exactly the image of Λ on L2 [0, ∞)

• Shows the map is invertible
Definition 8.
The inverse of the Laplace transform, Λ−1 : H2 → L2 [0, ∞) is
Z ∞
−1 1
u(t) = (Λ û)(t) = eσt · eıωt û(σ + ıω)dω
2π −∞
where σ can be any real number.

Corollary
Lemma 9.
hΛu, ΛyiH2 = hu, yiL2
• Thus Λ is unitary.
• L2 [0, ∞) and H2 are isomorphic.
kΛukH2
kΛk = sup =???
u∈L2 kukL2

H∞ - A Space of Bounded Analytic Functions
Definition 10.
A function Ĝ : C̄+ → Cn×m is in H∞ if
1. Ĝ(s) is analytic on the CRHP, C+ .
2.
lim Ĝ(σ + ıω) = Ĝ(ıω)
σ→0+
3.
sup σ̄(Ĝ(s)) < ∞
s∈C+
• A Banach Space with norm
kĜkH∞ = ess sup σ̄(Ĝ(ıω))

ω∈R

H∞ (A Signal Space) and Multiplier Operators
Every element of H∞ defines a multiplication operator.
Definition 11.
Given Ĝ ∈ H∞ , define MĜ ∈ L(H∞ )
(MĜ û)(s) = Ĝ(s)û(s)
for û ∈ H2 .
Functions vs. Operators

• Ĝ is a function of a complex variable.
• MĜ is an operator (a function of functions...).

Causal LTI Systems map to H∞
For any analytic functions, û and Ĝ, the function
ŷ(s) = Ĝ(s)û(s)
is analytic.
• Thus MĜ : H2 → H2 .
• Thus Λ−1 MĜ Λ maps L2 [0, ∞) → L2 [0, ∞).
Theorem 12.
G is a Causal, Linear, Time-Invariant Operator on L2 if and only if there exists
some Ĝ ∈ H∞ such that G = Λ−1 MĜ Λ.
(ΛGu)(ıω) = Ĝ(ıω)û(ıω)
H∞ is the space of transfer functions for linear time-invariant systems.

H∞ - The space of “Transfer Functions”
From Paley-Wiener, if G = Λ−1 MĜ Λ
Theorem 13.
kGkL(L2 ) = kMĜ kL(H2 ) = kĜkH∞
The Gain of the system G can be calculated as kĜkH∞

• This is the motivation for H∞ control
kGukL2
• minimize supu kukL2 .
I minimize maximum energy of the output.
Conclusion: H∞ provides a complete parametrization of the space of causal

bounded linear time-invariant operators.

Rational Transfer Functions (RH∞ )
The space of bounded analytic functions, H∞ is infinite-dimensional.
• this makes it hard to design optimal controllers.
We usually restrict ourselves to state-space systems and state-space controllers.
Definition 14.
The space of rational functions is defined as

p(s)
R := : p, q are polynomials
q(s)
We define the following rational subspaces.
RH2 = R ∩ H2
RH∞ = R ∩ H∞
Note that RH2 and RH∞ are not complete(Banach) spaces.

Rational Transfer Functions (RH∞ )
RH∞ is the set of proper rational functions with no poles in the closed right
half-plane (CRHP).
Definition 15.
p(s)
• A rational function r(s) = q(s) is Proper if the degree of p is less than or
equal to the degree of q.
p(s)
• A rational function r(s) = q(s) is Strictly Proper if the degree of p is less
than the degree of q.
Proposition 1.
1. Ĝ ∈ RH∞ if and only if Ĝ is proper with no poles on the closed right
half-plane.

State-Space Systems
Define a State-Space System G : L2 → L2 by y = Gu if
ẋ(t) = Ax(t) + Bu(t)
y(t) = Cx(t) + Du(t).
Theorem 16.
• For any stable state-space system, G, there exists some Ĝ ∈ RH∞ such
that
G = Λ−1 MĜ Λ
• For any Ĝ ∈ RH∞ , the operator G = Λ−1 MĜ Λ can be represented in
state-space for some A, B, C and D where A is Hurwitz.
For state-space system, (A, B, C, D),

Ĝ(s) = C(sI − A)−1 B + D
State-Space is NOT Unique. For any invertible T ,
• Ĝ = C(sI − A)−1 B + D = CT −1 (sI − T AT −1 )−1 T B + D.
I (A, B, C, D) and (T AT −1 , T B, CT −1 , D) both represent the system G.
Optimal Control Framework
tput
2-inputframework
2-output Framework
z regulated outputs exogenous inputs w

Plant
y sensed outputs actuator inputs u
We introduce the control framework by separating internal signals from external

signals.
nputs
Output are those inputs to the system that can be mani
u Signals:
• z: Output to be controlled/minimized
I Regulated output
• y: Output used by the controller
s inputs wMeasured
Iare allin real-time
other byinputs.
sensor
The same signal may appear in both outputs.
• e.g. if you can measure what you want to minimize.
outputs z are every output signal from the model.

2-input 2-output Framework
tput framework
z regulated outputs exogenous inputs w

Plant
y sensed outputs actuator inputs u
Input Signals:
• w: Disturbance, Tracking Signal, etc.
nputs u exogenous
are those
I input inputs to the system that can be mani
• u: Output from controller
I Input to actuator
I Not related to external input
s inputs w are all other inputs.

ngThe Optimal Control
interconnection is calledFramework
the (lower) star-product of P and K, or t
ional transformation (LFT).
The controller closes the loop from y to u.
z w
P
y u
romForwa to z is
linear given
system P , by
we have 4 subsystems.

P12 −w P K)−1 P
S(P, K) = Pz11 =+ PP1112K(I 22 21
y P21 P22 u
P11 : w 7→ z P12 : u 7→ z
P21 : w 7→ y P22 : u 7→ y
Note that all Pij can themselves be MIMO.
12 - 4 LFTs and stability
The RegulatorExample: the regulator

u nproc yp nsensor The plant P is given
   
z1 P0 0 P0
K + P0 +
z2 =  0 0 1 
12 - 4 LFTs and stability y P0 1 P20010
Example: the regulator Suppose P0 is
If we define q = w1Formulate
+ u andthe
r=above as
u then
P0 q, The plant Pẋ is

n y n
=given
proc
Ax +byB p sensor
   
z1z = y yp
1 p
P w1
n =w η
proc z1 P P0 + D
r 0=0 Cx w proc 1
z = z ==u w P= n == w
ηsensorz2 =  0 0 1  w
K + +
2
z2 u w2 0 sensor 2
ySubstituting
P0 1 P0 u
y = y1 = r + w2 + r P q
+
0
Suppose u
z2 P=0 is q = w1
Formulate the above as
zẋ1 = r +yBq
= Ax =r+
z1 = yp nproc = w1
z2 = u
P
nsensor = w2 leadsr =
to Cx + Dq

Substituting A B 0
+ P0 +
y r K q u 
z2 =Pu = q=Cw1D+ 0w2

z1 = r y =0r +0 w02
C D I
leads to
 
A B 0 B
y K u C D 0 D

P = 
0 0 0 I 
M. Peet Lecture 01: C D I /D
20 135
z1 P0 0 P0
K + P0 +
z2 =  0 0 1 
The Regulator y P0 1 P0
Suppose P0 is
Formulate the above as
ẋ = Ax + Bq
z1 = yp nproc = w1
P r = Cx + Dq
z2 = u nsensor = w2
Substituting
+ P0 +
r q
z2 = u q = w1 +
z1 = r y =r+w
leads to

A B 0 B
y K u C D 0 D

P =
0 0 0 I
C D I D
The reconfigured plant P is given by If P0 = (A, B, C, D), then
      
z1 (t) P0 0 P0 w1 (t) A B 0 B
z2 (t) =  0 0 I  w2 (t)  C D 0 D 
y(t) P0 I P0 u(t) P =  0 0 0 I 

C D I D

Diagnostics
2001.11.07.04
2001.11.07.04
Command inputs and diagnostic outputs

Command inputs and diagnostic outputs
zsystem w
zsystem wsystemsystem
System
System
yysystem usystem
usystem
system
Controller
Controller
commandinputs
command inputs diagnostic outputs
diagnostic outputs
Formulate
Formulate thetheabove
aboveasas
Plant
Plant
zsystem wsystem
zsystem
zdiag System wsystem
wcommands
zdiag System wcommands
ysystem usystem
yysystem Controller
commands udiag usystem
Controller
ycommands udiag

FTs and stability 200
Tracking Control
mple: a tracking problem
u nproc nsensor
−
e
r
K + P0 +
Define q = nproc P
+ u, then r = w1

−
z1 = e z = z1Werr= e −proc
e = rW P0 q nproc = w2
z2 u
   
z2 = u w1 r nsensor = w3
y Wact r W sens
y= 1 = w = w2  =  nproc 
y2 nsensor + P0 q
+ P0 + w3 nsensor
e = tracking error r = tracking input
nproc = process noise nsensor = sensor noise

r
K + P0 +
Tracking Control
P r = w1
−
z1 = e Werr Wproc nproc = w2
z2 = u Wact Wsens nsensor = w3
+ P0 +
y K u
  z1 = r − P0 (nproc + u)
I −P0 0 −P0
0 0 0 I  z2 = u
P =
I

0 0 0  y1 = r
0 P0 I P0 y2 = w3 + P0 (nproc + u)
ransformation (LFT).
Linear Fractional Transformation
Close the loop
z w
P
y u
toPlant:
z isgiven
by A

B1 B2

z P P12 w
= 11 where D12 
P =  C1 D11 −1
S(P, K) = P11 + P12K(I − P K) P
y P21 P22 u
C22
2 D21 D22 21

Controller: AK BK
u = Ky where K=
CK DK
Linear Fractional Transformation
z = P11 w + P12 u
y = P21 w + P22 u
u = Ky
Solving for u,
u = KP21 w + KP22 u
Thus
(I − KP22 )u = KP21 w
u = (I − KP22 )−1 KP21 w
Now we solve for z:

z = P11 + P12 (I − KP22 )−1 KP21 w

AK BK
LinearuFractional where
= Ky TransformationK̂ =
C K DK
This expression is called the Linear Fractional Transformation of (P, K), denoted
rconnection is S(P,called the (lower) star-product of P and
K) := P11 + P12 (I − KP22 )−1 KP21
AKA: Lower Star Product
z w
P
y u
to z is given by
AK BK
= Ky where K̂ =
Other Fractional Transformations
C K DK K
on is called the (lower) star-product

UpperofLFT
P and K, or the (lower)
tion (LFT).
Lower LFT: Upper LFT:
S(P, K) = P22 + P21Q(I − P11Q)−1P
z w
P Q
y u
z w
K P
y u
ven by
S(P, K) := P11 +P12 (I−KP22 )−1 KP21
S̄(P, K) := P22 +P21 Q(I −P11 K)−1 P12
P, K) = P11 + P12K(I − P22K)−1 P21

Other Fractional Transformations
Star Product
Star Product:
21 z1 w1
P
y u
K
z2 w2

K11 ) K) = P12 (I − K11 P22 )−1 K12
S(P,S(P,
S(P, K) :=
K21 (I − P22 K11 )−1 P21 S̄(K, P22 )
12 S(P, K11) P12(I − K11P22)−1K12
M. Peet −1 Lecture 01: 29 / 135
Well-Posedness
The interconnection doesn’t always make sense. Suppose
 
A B1 B2
AK BK
P =  C1 D11 D12  and K= .
CK DK
C2 D21 D22
Definition 17.
The interconnection S(P, K) is well-posed if for any smooth w and any x(0)
and xK (0), there exist functions x, xK , u, y, z such that
ẋ(t) = Ax(t) + B1 w(t) + B2 u(t) ẋK (t) = AK x(t) + BK y(t)

z(t) = C1 x(t) + D11 w(t) + D12 u(t) u(t) = CK x(t) + DK y(t)
y(t) = C2 x(t) + D21 w(t) + D22 u(t)
Note: The solution does not need to be in L2 .

• Says nothing about stability.

Well-Posedness
In state-space format, the closed-loop system is:

ẋ(t) A 0 x(t) B2 0 u(t) B1
= + + w(t)
ẋK (t) 0 AK xK (t) 0 BK y(t) 0

x(t) u(t)
z(t) = C1 0 + D12 0 + D11 w(t)
xK (t) y(t)
From
u(t) = DK y(t) + CK xK (t)
y(t) = D22 u(t) + C2 x(t) + D21 w(t)
We have

I −DK u(t) 0 CK x(t) 0
= + w(t)
−D22 I y(t) C2 0 xK (t) D21
Because the rest is state-space,

the interconnection is well-posed if and only if
I −DK
the matrix is invertible.
−D22 I
Well-Posedness
Question: When is
I −DK
−D22 I
invertible?
Answer: 2x2 matrices have a closed-form inverse
−1
I −DK I + DK QD22 DK Q
=
−D22 I QD22 Q
where Q = (I − D22 DK )−1 .
Proposition 2.
The interconnection S(P, K) is well-posed if and only if (I − D22 DK ) is
invertible.
• Equivalently (I − DK D22 ) is invertible.

• Sufficient conditions: DK = 0 or D22 = 0.
• To optimize over K, we will need to enforce this constraint somehow.
Optimal Control
Definition 18.
The Optimal H∞ -Control Problem is
min kS(P, K)kH∞

K∈H∞
• This is the Optimal H∞ Dynamic-Output-Feedback Control Problem
Another class of optimal control problem:
Definition 19.
The Optimal H2 -Control Problem is
min kS(P, K)kH2 such that

K∈H∞
S(P, K) ∈ H∞ .

Optimal Control
Choose K to minimize
kP11 + P12 (I − KP22 )−1 KP21 kH∞

AK BK
Equivalently choose to minimize
CK DK
 
A 0

B2 0

I −DK
−1
0 CK
B1 + B2 DK QD21

 0 AK + 0 BK −D22 I C2 0 BK QD21 
 
 −1 
I −DK 0 CK
C1 0 + D12 0
−D22 I C2 0
D11 + D12 DK QD21
H∞
where Q = (I − D22 DK )−1 .

In either case, the problem is Nonlinear.

Optimal Control
There are several ways to address the problem of nonlinearity.
kP11 + P12 (I − KP22 )−1 KP21 kH∞
Variable Substitution: The easiest way to make the problem linear is by
declaring a new variable R := (I − KP22 )−1 K
The optimization problem becomes: Choose R to minimize

kP11 + P12 RP21 kH∞ 2001.11.07.04
Realizability
z + P11 w
y u
P12 K + P21
P22
R
M. Peet z + Lecture 01: w 35 / 135

Optimal Control
We optimize
kP11 + P12 (I − KP22 )−1 KP21 kH∞ = kP11 + P12 RP21 kH∞
Once, we have the optimal R, we can recover the optimal K as
K = R(I + RP22 )−1
Problems:
• how to optimize k·kH∞ .
• Is the controller stable?
I Does the inverse (I + RP −1
22 ) exist? Yes.
I Is it a bounded linear operator?
I In which space?
• An important branch of control.

I Coprime factorization
I Youla parameterization
• We will sidestep this body of work.

What is Optimization?
Optimization can be posed in functional form:
min objective function : subject to
x∈F
inequality constraints
which may have the form
min f0 (x) : subject to
x∈F
fi (x) ≥ 0 i = 1, · · · k
Special Cases:
• Linear Programming
I fi (x) = Ax − b ( Affine functions with fi : Rn → Rm )
I EASY: Simplex/Ellipsoid Algorithm
• Polynomial Programming
I The f : Rn → Rm are polynomials. (NP-HARD)
i
• Semidefinite Programming
I The fi : Rn → Rm×m are affine. (EASY)
For semidefinite programming, what does fi (x) ≥ 0 mean?

How Hard is Optimization?
Why is Linear Programming easy and polynomial programming hard?
min f0 (x) : subject to
x∈F
fi (x) ≥ 0 i = 1, · · · k
The Geometric Representation is equivalent:
min f0 (x) : subject to x∈S
x∈F
where S := {x : fi (x) ≥ 0, i = 1, · · · , k}.
The Pure Geometric Representation:

min γ : subject to
γ,x∈F
(γ, x) ∈ S 0
where S 0 := {(γ, x) : γ − f0 (x) ≥ 0, fi (x) ≥ 0, i = 1, · · · , k}.
• Two optimization problems are Equivalent if a solution to one can be used
to construct a solution to the other.
Convexity
Definition 20.
A set is convex if for any x, y ∈ Q,
{µx + (1 − µ)y : µ ∈ [0, 1]} ⊂ Q.
The line connecting any two points lies in the set.

subject to Ax = b
Cx ≤ d
Convex Optimization
Example
Convex Optimization:
minimize
x 1 + x2
Definition 21.subject to 3x1 + x2 ≥ 3
Consider the optimization problemx2 ≥ 1
x1 ≤ 4
min γ : subject
−x1 + 5xto 2 ≤ 20
γ,x∈F
0
x 1 + 4x 2 ≤ 20
(γ, x) ∈ S .
The problem is Convex Optimization if the set

S 0 is convex.
Convex optimization problems have the property that the Gradient projection
algorithm (or Newton iteration with barrier functions) will always converge to
the global optimal.
The question is, of course, when is the set S 0 convex?
• For polynomial optimization, a sufficient condition is that all functions fi
are convex.
I The level set of a convex function is a convex set.

Non-Convexity and Local Optima
Newton’s Algorithm: Designed to solve f (x∗ ) = 0 (is min f (x) ≥ 0?)
f (xk )
xk+1 = xk − t
f 0 (xk )
∗
where t is the step-size. (From df /dx ∼ f (x)−f (x )
= x−x∗ )
For non-convex optimization, Newton descent may get stuck at local optima.
For constrained optimization, constraints are represented by barrier functions.

Convex Cones
Definition 22.
A set is a cone if for any x ∈ Q,
{µx : µ ≥ 0} ⊂ Q.
A subspace is a cone but not all cones are subspaces.

• If the cone is also convex, it is a convex cone.
• Cones are convex if they are closed under addition.

What is an Inequality Constraint?
Question: What does f (x) ≥ 0 mean.

• What does y ≥ 0 mean?
If y is a Scalar (y ∈ R), then y ≥ 0 if y ∈ [0, ∞].
Question: What if y is a vector (y ∈ Rn )?

• Then we have several options...
Examples: Let y ∈ Rn .
• Positive Orthant: y ≥ 0 if yi ≥ 0 for i = 1, · · · , n.
P
• Half-space: y ≥ 0 if yi ≥ 0 (1T y ≥ 0).
I More generally, y ≥ 0 if aT y + b ≥ 0.
• Intersection of Half-spaces: y ≥ 0 if aT i y + bi ≥ 0 for i = 1, · · · , n.

I The positive orthant is the intersection of half-spaces with b = 0 and
i
ai = ei (unit vectors).
Question: What if y is a matrix???

Positivity
What is an inequality? What does ≥ 0 mean?

• An inequality implies a partial ordering:
I x ≥ y if x − y ≥ 0
• Any convex cone, C defines a partial ordering:
I x − y ≥ 0 if x − y ∈ C
• The ordering is only partial because x 6≤ 0 does not imply x ≥ 0

I −x 6∈ C does not imply x ∈ C.
I x may be indefinite.
Conclusion:
• Convex Optimization includes positivity induced from any partial ordering.
• In particular, we focus on Matrix Positivity.

Matrix Positivity
Definition 23.
A symmetric matrix P ∈ Sn is Positive Semidefinite, denoted P ≥ 0 if
xT P x ≥ 0 for all x ∈ Rn
Definition 24.
A symmetric matrix P ∈ Sn is Positive Definite, denoted P > 0 if
xT P x > 0 for all x 6= 0
• P is Negative Semidefinite if −P ≥ 0
• P is Negative Definite if −P > 0
• A matrix which is neither Positive nor Negative Semidefinite is Indefinite
The set of positive or negative matrices is a convex cone.

Positive Matrices
Lemma 25.
P ∈ Sn is positive definite if and only if all its eigenvalues are positive.
Things which are easy to prove:

• A Positive Definite matrix is invertible.
• The inverse of a positive definite matrix is positive definite.
• If P > 0, then T P T T ≥ 0 for any T . If T is invertible, then T P T T > 0.
Lemma 26.
1
For any P > 0, there exists a positive square root, P 2 > 0 such that
1 1
P = P 2P 2.

Semidefinite Programming - Dual Form
minimize trace CX
subject to trace Ai X = bi for all i
X0
• The variable X is a symmetric matrix

• X 0 means X is positive semidefinite
• The feasible set is the intersection of an affine set with the positive
semidefinite cone
X ∈ Sn | X 0
P
Recall trace CX = i,j Ci,j Xj,i .

SDPs with Explicit Variables - Primal Form
We can also explicitly parametrize the affine set to give
minimize cT x
subject to F0 + x1 F1 + x2 F2 + · · · + xn Fn 0
where F0 , F1 , . . . , Fn are symmetric matrices.
The inequality constraint is called a Linear Matrix Inequality (LMI); e.g.,

 
x1 − 3 x1 + x2 −1
x1 + x2 x2 − 4 0 0
−1 0 x1
which is equivalent to
     
−3 0 −1 1 1 0 0 1 0
 0 −4 0  + x1 1 0 0 + x2 1 1 0 0
−1 0 0 0 0 1 0 0 0

Linear Matrix Inequalities
Linear Matrix Inequalities are often a Simpler way to solve control problems.
Common Form:
Find X :
X
Ai XBi + Q > 0
i
The most important Linear Matrix Inequality is the Lyapunov Inequality.

There are several very efficient LMI/SDP Solvers for Matlab:
• SeDuMi
I Fast, but somewhat unreliable.
I See http://sedumi.ie.lehigh.edu/
• LMI Lab (Part of Matlab’s Robust Control Toolbox)
I Universally disliked
I See http://www.mathworks.com/help/robust/lmis.html
• YALMIP (a parser for other solvers)

I See http://users.isy.liu.se/johanl/yalmip/
I recommend YALMIP with solver SeDuMi.

Semidefinite Programming(SDP):
Common Examples in Control
Some Simple examples of LMI conditions in control include:

• Stability
AT X + XA ≺ 0
X0
• Stabilization
AX + BZ + XAT + Z T B T ≺ 0
X0
• H2 Synthesis
min T r(W )

X AT
A B2 + X ZT + B1 B1T ≺ 0
Z B2T

X (CX + DZ)T
0
(CX + DZ) W
We will go beyond these examples.
Lyapunov Theory
LMIs unite time-domain and frequency-domain analysis
ẋ(t) = f (x(t))
Theorem 27 (Lyapunov).
Suppose there exists a continuously differentiable function V for which V (0) = 0
and V (x) > 0 for x 6= 0. Furthermore, suppose limkxk→∞ V (x) = ∞ and
V (x(t + h)) − V (x(t)) d

lim+ = V (x(t)) < 0
h→0 h dt
for any x such that ẋ(t) = f (x(t)). Then for any x(0) ∈ R the system of
equations
ẋ(t) = f (x(t))
has a unique solution which is stable in the sense of Lyapunov.

The Lyapunov Inequality (Our First LMI)
Lemma 28.
A is Hurwitz if and only if there exists a P > 0 such that
AT P + P A < 0
Proof.
Suppose there exists a P > 0 such that AT P + P A < 0.
• Define the Lyapunov function V (x) = xT P x.
• Then V (x) > 0 for x 6= 0 and V (0) = 0.
• Furthermore,
V̇ (x(t)) = ẋ(t)T P x(t) + x(t)T P ẋ(t)
= x(t)T AT P x(t) + x(t)T P Ax(t)

= x(t)T AT P + P A x(t)
• Hence V̇ (x(t)) < 0 for all x 6= 0. Thus the system is globally stable.
• Global stability implies A is Hurwitz.
The Lyapunov Inequality
Proof.
For the other direction, if A is Hurwitz, let
Z ∞
T
P = eA s eAs ds
0
• Converges because A is Hurwitz.

Z ∞
• Furthermore T
PA = eA s eAs Ads
Z0 ∞ Z ∞
T T d As
= eA s AeAs ds = eA s e ds
0 0 ds
∞ Z ∞
T d AT s As
= eA s eAs − e e
0 0 ds
Z ∞
T
= −I − AT eA s eAs = −I − AT P
0
T
• Thus P A + A P = −I < 0.

The Lyapunov Inequality
Other Versions:
Lemma 29.
(A, B) is controllable if and only if there exists a X > 0 such that
AT X + XA + BB T ≤ 0
Lemma 30.
(C, A) is observable if and only if there exists a X > 0 such that
AX + XAT + C T C ≤ 0

The Static State-Feedback Problem
Lets start with the problem of stabilization.
Definition 31.
The Static State-Feedback Problem is to find a feedback matrix K such that
ẋ(t) = Ax(t) + Bu(t)

u(t) = Kx(t)
is stable
• Find K such that A + BK is Hurwitz.
Can also be put in LMI format:
Find X > 0, K :
X(A + BK) + (A + BK)T X < 0
Problem: Bilinear in K and X.

The Static State-Feedback Problem
• The bilinear problem in K and X is a common paradigm.
• Bilinear optimization is not convex.
• To convexify the problem, we use a change of variables.
Problem 1:
Find X > 0, K :
X(A + BK) + (A + BK)T X < 0
Problem 2:
Find P > 0, Z :
AP + BZ + P AT + Z T B T < 0
Definition 32.
Two optimization problems are equivalent if a solution to one will provide a
solution to the other.
Theorem 33.
Problem 1 is equivalent to Problem 2.
The Dual Lyapunov Equation
Problem 1: Problem 2:
Find X > 0, : Find Y > 0, :
XA + AT X < 0 Y AT + AY < 0
Lemma 34.
Problem 1 is equivalent to problem 2.
Proof.
First we show 1) solves 2). Suppose X > 0 is a solution to Problem 1. Let
Y = X −1 > 0.
• If XA + AT X < 0, then
X −1 (XA + AT X)X −1 < 0

• Hence
X −1 (XA + AT X)X −1 = AX −1 + X −1 AT = AY + Y AT < 0
• Therefore, Problem 2 is feasible with solution Y = X −1 .

The Dual Lyapunov Equation
Find X > 0, : Find Y > 0, :
XA + AT X < 0 Y AT + AY < 0
Proof.
Now we show 2) solves 1) in a similar manner. Suppose Y > 0 is a solution to
Problem 1. Let X = Y −1 > 0.
• Then
XA + AT X = X(AX −1 + X −1 AT )X
= X(AY + Y AT )X < 0
Conclusion: If V (x) = xT P x proves stability of ẋ = Ax,

• Then V (x) = xT P −1 x proves stability of ẋ = AT x.

The Stabilization Problem
Thus we rephrase Problem 1
Find P > 0, K : Find X > 0, Z :

T
(A + BK)P + P (A + BK) < 0 AX + BZ + XAT + Z T B T < 0
Theorem 35.
Problem 1 is equivalent to Problem 2.
Proof.
We will show that 2) Solves 1). Suppose X > 0, Z solves 2). Let P = X > 0
and K = ZP −1 . Then Z = KP and
(A + BK)P + P (A + BK)T = AP + P AT + BKP + P K T B T

= AP + P AT + BZ + Z T B T < 0
Now suppose that P > 0 and K solve 1). Let X = P > 0 and Z = KP . Then
AP + P AT + BZ + Z T B T = (A + BK)P + P (A + BK)T < 0
The Stabilization Problem
The result can be summarized more succinctly
Theorem 36.
(A, B) is static-state-feedback stabilizable if and only if there exists some P > 0
and Z such that
AP + P AT + BZ + Z T B T < 0
with u(t) = ZP −1 x(t).
Standard Format:

P AT
T
A B + P Z <0
Z BT

The Schur complement
Before we get to the main result, recall the Schur complement.
Theorem 37 (Schur Complement).

For any S ∈ Sn , Q ∈ Sm and R ∈ Rn×m , the following are equivalent.

M R
1. >0
RT Q
2. Q > 0 and M − RQ−1 RT > 0
A commonly used property of positive matrices.

Also Recall: If X > 0,
• then X − I > 0 for sufficiently small.

The KYP Lemma (AKA: The Bounded Real Lemma)
The most important theorem in this lecture.
Lemma 38 (KYP Lemma).
Suppose
A B
Ĝ(s) = .
C D
Then the following are equivalent.
• kGkH∞ ≤ γ.
• There exists a X > 0 such that
T
A X + XA XB 1 CT
+ C D <0
BT X −γI γ DT
Can be used to calculate the H∞ -norm of a system

• Originally used to solve LMI’s using graphs. (Before Computers)
• Now used directly instead of graphical methods like Bode.
The feasibility constraints are linear
• Can be combined with other methods.
The KYP Lemma
Proof.
We will only show that ii) implies i). The other direction requires the
Hamiltonian, which we have not discussed.
• We will show that if y = Gu, then kykL2 ≤ γkukL2 .
• From the 1 x 1 block of the LMI, we know that AT X + XA < 0, which
means A is Hurwitz.
• Because the inequality is strict, there exists some > 0 such that
T
A X + XA XB 1 CT
T + T C D
B X −(γ − )I γ D
T T
A X + XA XB 1 C 0 0
= + C D + <0
BT X −γI γ DT 0 I
• Let y = Gu. Then the state-space representation is
y(t) = Cx(t) + Du(t)

ẋ(t) = Ax(t) + Bu(t) x(0) = 0
The KYP Lemma
Proof.
• Let V (x) = xT Xx. Then the LMI implies
T " T #
x(t) A X + XA XB 1 CT x(t)
+ C D
u(t) BT X −(γ − )I γ DT u(t)
T T T T
x A X + XA XB x 1 x C x
= + C D
u BT X −(γ − )I u γ u DT u
T T
x A X + XA XB x 1
= + yT y
u BT X −(γ − )I u γ
1
= xT (AT X + XA)x + xT XBu + uT B T Xx − (γ − )uT u + y T y
γ
T T T 1 T
= (Ax + Bu) Xx + x X(Ax + Bu) − (γ − )u u + y y
γ
1
= ẋ(t)T Xx(t) + x(t)T X ẋ(t) − (γ − )ku(t)k2 + ky(t)k2
γ
2 1 2
= V̇ (x(t)) − (γ − )ku(t)k + ky(t)k < 0
γ
The KYP Lemma
Proof.
1
• Now we have V̇ (x(t)) − (γ − )ku(t)k2 + ky(t)k2 < 0
γ
• Integrating in time, we get
Z T
1
V̇ (x(t)) − (γ − )ku(t)k2 + ky(t)k2 dt
0 γ
Z T Z
2 1 T
= V (x(T )) − V (x(0)) − (γ − ) ku(t)k dt + ky(t)k2 dt < 0
0 γ 0
• Because A is Hurwitz, limT →∞ x(T ) = 0.

• Hence limT →∞ V (x(T )) = 0.
• Likewise, because x(0) = 0, we have V (x(0)) = 0.

The KYP Lemma
Proof.
• Since V (x(0)) = V (x(∞)) = 0,
Z T Z
2 1 T 2
lim V (x(T )) − V (x(0)) − (γ − ) ku(t)k dt + ky(t)k dt
T →∞ γ 0
Z ∞ Z0 ∞
1
= 0 − 0 − (γ − ) ku(t)k2 dt + ky(t)k2 dt
0 γ 0
1
= −(γ − )kuk2L2 + kyk2L2 dt < 0
γ
• Thus
kyk2L2 dt < (γ 2 − γ)kuk2L2
• By definition, this means kGk2H∞ ≤ (γ 2 − γ) < γ 2 or
kGkH∞ < γ

The Positive Real Lemma
A Passivity Condition
A Variation on the KYP lemma is the positive-real lemma
Lemma 39.
Suppose
A B
Ĝ(s) = .
C D
• G is passive. i.e. (hu, GuiL2 ≥ 0).
• There exists a P > 0 such that
T
A P + P A P B − CT
≤0
B T P − C −DT − D

ransformation (LFT). Transformation
Recall: Linear Fractional
z w
P
y u
toPlant:
z isgiven
by A

B1 B2

z P P12 w
= 11 where P =  C1 D11 −1
D12 
S(P, K) = P11 + P12K(I − P K) P
y P21 P22 u
C22
2 D21 D22 21

Controller: AK BK
u = Ky where K=
CK DK
Optimal Control
kP11 + P12 (I − KP22 )−1 KP21 k

AK BK
CK DK
 
A 0

B2 0

I −DK
−1
0 CK
B1 + B2 DK QD21

 0 AK + 0 BK −D22 I C2 0 BK QD21 
 
 −1 
I −DK 0 CK
C1 0 + D12 0
−D22 I C2 0
D11 + D12 DK QD21
H∞
where Q = (I − D22 DK )−1 .

The following interconnection
Optimal Full-State isFeedback
called the (lower) star-product of P and K, or the (lower)
Control
linear-fractional transformation (LFT).
z w
P
y u
The map from w to z is given by

For the full-state feedback case, we consider a controller of the form
S(P, K) = P11 + P12K(I − P22K)−1 P21
u(t) = F x(t)

Controller: 0 0
u = Ky where K=
0 F
Plant:  
A B1 B2
z P P12 w
= 11 where P =  C1 D11 D12 
y P21 P22 u
I 0 0

Optimal Full-State Feedback Control
Thus the closed-loop state-space representation is

A + B2 F B1
S(P̂ , K̂) =
C1 + D12 F D11
By the KYP lemma, kS(P̂ , K̂)kH∞ < γ if and only if there exists some X > 0
such that

(A + B2 F )T X + X(A + B2 F ) XB1
B1T X −γI

1 (C1 + D12 F )T
+ T (C1 + D12 F ) D11 < 0
γ D11
This is a matrix inequality, but is nonlinear

• Quadratic (Not Bilinear)
• May NOT apply variable substitution trick.

Schur Complement
The KYP condition is
T
A X + XA XB 1 CT
+ C D <0
BT X −γI γ DT
Recall the Schur Complement
Theorem 40 (Schur Complement).

For any S ∈ Sn , Q ∈ Sm and R ∈ Rn×m , the following are equivalent.

M R
1. <0
RT Q
2. Q < 0 and M − RQ−1 RT < 0
In this case, let Q = − γ1 I < 0,

T
A X + XA XB T
M= R= C D
BT X −γI
Note we are making the LMI Larger.

Schur Complement
The Schur Complement says that

T
A X + XA XB 1 CT
+ C D <0
BT X −γI γ DT
if and only if  T 
A X + XA XB CT
 BT X −γI DT  < 0
C D −γI
This leads to the
Full-State Feedback Condition
 
(A + B2 F )T X + X(A + B2 F ) XB1 (C1 + D12 F )T
 B1T X −γI T
D11 <0
(C1 + D12 F ) D11 −γI
which is now bilinear in X and F .

Dual KYP Lemma
To apply the variable substitution trick, we must also construct the dual form of
this LMI.
Lemma 41 (KYP Dual).
Suppose
A B
Ĝ(s) = .
C D
• kGkH∞ ≤ γ.
• There exists a Y > 0 such that
 
Y AT + AY B Y CT
 BT −γI DT  < 0
CY D −γI

Dual KYP Lemma
Proof.
Let X = Y −1. Then 
Y AT + AY XB Y CT
 BT X −γI DT  < 0 and Y >0
CY D −γI
if and only
 −1 if X > 0 and
   −1 
Y 0 0 Y AT + AY B Y CT Y 0 0
 0 I 0  BT −γI DT   0 I 0
0 0 I CY D −γI 0 0 I
 T T

A X + XA XB C
=  BT X −γI DT  < 0.
C D −γI
By the Schur complement

T this is equivalent
to
A X + XA XB 1 CT
+ C D <0
BT X −γI γ DT
By the KYP lemma, this is equivalent to kGkH∞ ≤ γ.

Full-State Feedback Optimal Control
We can now apply this result to the state-feedback problem.
Theorem 42.
The following are equivalent:
• There exists an F such that kS(P, K(0, 0, 0, F ))kH∞ ≤ γ.
• There exist Y > 0 and Z such that
 
Y AT + AY + Z T B2T + B2 Z B1 Y C1T + Z T D12
T
 B1T −γI T
D11 <0
C1 Y + D12 Z D11 −γI
One may use F = ZY −1 .

Proof.
Suppose there exists an F such that kS(P, K(0, 0, 0, F ))kH∞ ≤ γ. By the Dual
KYP lemma, this implies there exists a Y > 0 such that
 
Y (A + B2 F )T + (A + B2 F )Y B1 Y (C1 + D12 F )T
 B1T −γI T
D11 <0
(C1 + D12 F )Y D11 −γI
Let Z = F Y . Then
 
Y AT + Z T B2T + AY + B2 Z B1 Y C1T + Z T D12 T T
)
 B1T −γI T
D11 
C1 Y + D12 Z D11 −γI
 
Y AT + Y F T B2T + AY + B2 F Y B1 Y C1T + Y F T D12
T T
)
= B1T −γI D11T 
C1 Y + D12 F Y D11 −γI
 
= B1T −γI T
D11  < 0.
(C1 + D12 F )Y D11 −γI

Proof.
Now suppose there exists a Y > 0 and Z such that
 
Y AT + Z T B2T + AY + B2 Z B1 Y C1T + Z T D12
T
 B1T −γI T
D11 <0
C1 Y + D12 Z D11 −γI
Let F = ZY −1 . Then
 
 B1T −γI T
D11 
(C1 + D12 F )Y D11 −γI
 T T T

Y A + Y F B2 + AY + B2 F Y B1 Y C1T + Y F T D12
T
= B1T −γI T
D11 
C1 Y + D12 F Y D11 −γI
 T T T T T T

Y A + Z B2 + AY + B2 Z B1 Y C1 + Z D12
= B1T −γI T
D11 <0
C1 Y + D12 Z D11 −γI
Therefore the following optimization problems are equivalent

Form A
minkS(P, K(0, 0, 0, F ))kH∞

F
Form B
min γ :
γ,Y,Z
 
−Y 0 0 0
 0 Y A T
+ AY + Z T B2T + B2 Z B1 T 
Y C1T + Z T D12
 <0
 0 B1T −γI T
D11 
0 C1 Y + D12 Z D11 −γI
The optimal controller is given by F = ZY −1 .

Next: Optimal Output Feedback

rconnection is called the (lower) star-product of P and
Optimal Output Feedback
Recall: Linear Fractional Transformation
z w
P
y u
toPlant:
z isgiven
by A

B1 B2

z P P12 w
= 11 where P =  C1 D11 D12 
y P21 P22 u −1
S(P, K) = P11 + P12K(I − P K) P C22
2 D21 D22 21

Controller: AK BK
u = Ky where K=
CK DK
Optimal Control
kP11 + P12 (I − KP22 )−1 KP21 k

AK BK
CK DK
 
A 0

B2 0

I −DK
−1
0 CK
B1 + B2 DK QD21

 0 AK + 0 BK −D22 I C2 0 BK QD21 
 
 −1 
I −DK 0 CK
C1 0 + D12 0
−D22 I C2 0
D11 + D12 DK QD21
H∞
where Q = (I − D22 DK )−1 .

Optimal Control
Recall that −1
I −DK I + DK QD22 DK Q
=
−D22 I QD22 Q
where Q = (I − D22 DK )−1 . Then
−1
A 0 B2 0 I −DK 0 CK
Acl := +
0 AK 0 BK −D22 I C2 0

A 0 B2 0 I + DK QD22 DK Q 0 CK
= +
0 AK 0 BK QD22 Q C2 0

A + B2 DK QC2 B2 (I + DK QD22 )CK
=
BK QC2 AK + BK QD22 CK
Likewise

I + DK QD22 DK Q 0 CK
Ccl := C1 0 + D12 0
QD22 Q C2 0

= C1 + D12 DK QC2 D12 (I + DK QD22 )CK

Optimal Output Feedback Control
Thus we have
 
A + B2 DK QC2 B2 (I + DK QD22 )CK B1 + B2 DK QD21
 BK QC2 AK + BK QD22 CK BK QD21 

C1 + D12 DK QC2 D12 (I + DK QD22 )CK D11 + D12 DK QD21
where Q = (I − D22 DK )−1 .

• This is nonlinear in (AK , BK , CK , DK ).
• Hence we make a change of variables (First of several).
AK2 = AK + BK QD22 CK
BK2 = BK Q
CK2 = (I + DK QD22 )CK
DK2 = DK Q

This yields the system

 
A + B2 DK2 C2 B2 CK2 B1 + B2 DK2 D21
 BK2 C2 AK2 BK2 D21 

C1 + D12 DK2 C2 D12 CK2 D11 + D12 DK2 D21

AK2 BK2
Which is affine in .
CK2 DK2

Hence we can optimize over our new variables.
• However, the change of variables must be invertible.
If we recall that
(I − QM )−1 = I + Q(I − M Q)−1 M
then we get
I + DK QD22 = I + DK (I − D22 DK )−1 D22 = (I − DK D22 )−1
Examine the variable CK2
CK2 = (I + DK (I − D22 DK )−1 D22 )CK

= (I − DK D22 )−1 CK
Hence, given CK2 , we can recover CK as
CK = (I − DK D22 )CK2

Now suppose we have DK2 . Then
DK2 = DK Q = DK (I − D22 DK )−1
implies that
DK = DK2 (I − D22 DK ) = DK2 − DK2 D22 DK
or
(I + DK2 D22 )DK = DK2
which can be inverted to get
DK = (I + DK2 D22 )−1 DK2

Once we have CK and DK , the other variables are easily recovered as
BK = BK2 Q−1 = BK2 (I − D22 DK )

AK = AK2 − BK (I − D22 DK )−1 D22 CK
To summarize, the original variables can be recovered as
DK = (I + DK2 D22 )−1 DK2

BK = BK2 (I − D22 DK )
CK = (I − DK D22 )CK2
AK = AK2 − BK (I − D22 DK )−1 D22 CK

 
A + B2 DK2 C2 B2 CK2 B1 + B2 DK2 D21
Acl Bcl
:=  BK2 C2 AK2 BK2 D21 
Ccl Dcl
C1 + D12 DK2 C2 D12 CK2 D11 + D12 DK2 D21
   
A 0 B1 0 B2
Acl Bcl A BK2 0 I 0
=0 0 0  + I 0  K2
Ccl Dcl CK2 DK2 C2 0 D21
C1 0 D11 0 D12
Or

A 0 0 B2 AK2 BK2 0 I
Acl = +
0 0 I 0 CK2 DK2 C2 0

B1 0 B2 AK2 BK2 0
Bcl = +
0 I 0 CK2 DK2 D21

AK2 BK2 0 I
Ccl = C1 0 + 0 D12
CK2 DK2 C2 0

AK2 BK2 0
Dcl = D11 + 0 D12
CK2 DK2 D21
Lemma 43 (Transformation Lemma).

Suppose that
Y1 I
>0
I X1
Then there exist X2 , X3 , Y2 , Y3 such that
−1
X1 X2 Y1 Y2
X= = = Y −1 > 0
X2T X3 Y2T Y3

Y1 I
where Ycl = has full rank.
Y2T 0

Transformation Lemma
Proof.
• Since
Y1 I
> 0,
I X1
by the Schur complement X1 > 0 and X1−1 − Y1 > 0. Since
I − X1 Y1 = X1 (X1−1 − Y1 ), we conclude that I − X1 Y1 is invertible.
• Choose any two square invertible matrices X2 and Y2 such that
X2 Y2T = I − X1 Y1
• Because X2 and Y2 are non-singular,

Y1 Y2 I 0
YclT = and Xcl =
I 0 X1 X2
are also non-singular.

Transformation Lemma
Proof.
• Now define X and Y as
X = Ycl−T Xcl and −1 T

Y = Xcl Ycl .
Then
XY = Ycl−1 Xcl Xcl
−1
Ycl = I
Likewise, Y X = I. Hence, Y = X −1 .

Lemma 44 (Converse Transformation Lemma).

X1 X2
Given X = > 0 where X2 has full column rank. Let
X2T X3

−1 Y1 Y2
X =Y =
Y2T Y3
then
Y1 I
>0
I X1

Y1 I
and Ycl = has full column rank.
Y2T 0

Converse Transformation Lemma
Proof.
I 0
Since X2 is full rank, Xcl = also has full column rank. Note that
X1 X2
XY = I implies

Y Y2 X1 X2 I 0
YclT X = 1 = = Xcl .
I 0 X2T X3 X1 X2
Hence
Y Y2 I 0
YclT = 1 = Y = Xcl Y
I 0 X1 X2
has full column rank. Now, since XY = I implies X1 Y1 + X2 Y2T = I, we have

I 0 Y1 I Y1 I Y1 I
Xcl Ycl = = =
X1 X2 Y2T 0 X1 Y1 + X2 Y2T X1 I X1
Furthermore, because Ycl has full rank,

Y1 I T
= Xcl Ycl = Xcl Y Xcl = YclT XYcl > 0
I X1
Theorem 45.
The following are equivalent.

AK BK
• There exists a K̂ = such that kS(P, K)kH∞ < γ.
CK DK

X1 I
• There exist X1 , Y1 , An , Bn , Cn , Dn such that >0
I Y1
 
AY1 +Y1 AT +B2 Cn +CnT B2T ∗T ∗T ∗T
T T T T T T
 A + An + [B2 Dn C2 ] X1 A+A X1 +Bn C2 +C2 Bn ∗ ∗T 
 [B1 + B2 Dn D21 ] T
[XB1 + Bn D21 ] T
−γI
<0
C1 Y1 + D12 Cn C1 + D12 Dn C2 D11 +D12 Dn D21 −γI
Moreover,
−1 −1
AK2 BK2 X X1 B2 An Bn X1 AY1 0 Y2T 0
= 2 −
CK2 DK2 0 I Cn Dn 0 0 C2 Y1 I
for any full-rank X2 and Y2 such that
−1
X1 X2 Y1 Y2
=
X2T X3 Y2T Y3

Proof: If.
Suppose there exist X1 , Y1 , An , Bn , Cn , Dn such that the LMI is feasible. Since

X1 I
> 0,
I Y1
by the transformation lemma, there exist X2 , X3 , Y2 , Y3 such that

−1
X X2 Y Y2
X := = >0
X2T X3 Y2T Y3

Y I AK BK
where Ycl = has full row rank. Let K = where
Y2T 0 CK DK
DK = (I + DK2 D22 )−1 DK2

BK = BK2 (I − D22 DK )
CK = (I − DK D22 )CK2
AK = AK2 − BK (I − D22 DK )−1 D22 CK .
Proof: If.
and where
−1 T −1
AK2 BK2 X2 X1 B2 An Bn X1 AY1 0 Y2 0
= − .

Proof: If.
As discussed previously, this means the closed-loop system is
   
A 0 B1 0 B2
Acl Bcl A BK2 0 I 0
=0 0 0  + I 0  K2
Ccl Dcl CK2 DK2 C2 0 D21
C1 0 D11 0 D12
   
A 0 B1 0 B2
=0 0 0  + I 0 
C1 0 D11 0 D12
−1 T −1
X2 X1 B2 An B n X1 AY1 0 Y2 0 0 I 0
−
0 I Cn Dn 0 0 C2 Y1 I C2 0 D21
Now look at the LMI from the KYP lemma.

Proof: If.
Expanding out, we obtain
 T  T T
 
Ycl 0 0 Acl X + XAcl XBcl Ccl Ycl 0 0
 0 I 0  T
Bcl X −γI T 
Dcl 0 I 0 =
0 0 I Ccl Dcl −γI 0 0 I
 
T T T T T T
 [B1 + B2 Dn D21 ] T
[XB1 + Bn D21 ]T
−γI
<0

Acl Bcl
Hence, by the KYP lemma, S(P, K) = satisfies
Ccl Dcl
kS(P, K)kH∞ < γ.

Proof: Only If.

AK BK
Now suppose that kS(P, K)kH∞ < γ for some K = . Since
CK DK
kS(P, K)kH∞ < γ, by the KYP lemma, there exists a

X1 X2
X= >0
X2T X3
such that  T 
T
Acl X + XAcl XBcl Ccl
 T
Bcl X −γI T 
Dcl <0
Ccl Dcl −γI
Because the inequalities are strict, we can assume that X2 has full row rank.
Define
Y1 Y2 −1 Y1 I
Y = =X and Ycl =
Y2T Y3 Y2T 0
Then, according to the converse transformation
lemma, Ycl has full row rank
and X1 I
> 0.
I Y1
Proof: Only If.
Now, using the given AK , BK , CK , DK , define the variables
T
An B n X2 X1 B2 AK2 BK2 Y2 0 X1 AY1 0
= + .
Cn Dn 0 I CK2 DK2 C2 Y1 I 0 0
where
AK2 = AK + BK (I − D22 DK )−1 D22 CK BK2 = BK (I − D22 DK )−1
CK2 = (I + DK (I − D22 DK )−1 D22 )CK DK2 = DK (I − D22 DK )−1
Then as before
   
A 0 B1 0 B2
Acl Bcl
=0 0 0  + I 0 
Ccl Dcl
C1 0 D11 0 D12
−1 T −1
X2 X1 B2 An Bn X1 AY1 0 Y2 0 0 I 0
−
0 I Cn Dn 0 0 C2 Y1 I C2 0 D21

Proof: Only If.

Expanding out the LMI, we find
 T T T 
AY1 +Y1 A +B2 Cn +Cn B2 ∗T ∗T ∗T
 AT + An + [B2 Dn C2 ]T X1 A+AT X1 +Bn C2 +C2T BnT ∗T ∗T 
 [B1 + B2 Dn D21 ] T
[XB1 + Bn D21 ] T
−γI

 T  T T
 
Ycl 0 0 Acl X + XAcl XBcl Ccl Ycl 0 0
= 0 I 0  Bcl T
Xcl −γI T 
Dcl 0 I 0 < 0
0 0 I Ccl Dcl −γI 0 0 I

Conclusion
To solve the H∞ -optimal state-feedback problem, we solve
min γ such that

γ,X1 ,Y1 ,An ,Bn ,Cn ,Dn

X1 I
>0
I Y1
 
T T T T T T
 [B1 + B2 Dn D21 ]T [XB1 + Bn D21 ]T −γI
< 0

Conclusion
Then, we construct our controller using
DK = (I + DK2 D22 )−1 DK2

BK = BK2 (I − D22 DK )
CK = (I − DK D22 )CK2
AK = AK2 − BK (I − D22 DK )−1 D22 CK .
where
−1 T −1
AK2 BK2 X2 X1 B2 An Bn X1 AY1 0 Y2 0
= − .
and where X2 and Y2 are any matrices which satisfy X2 Y2 = I − X1 Y1 .

• e.g. Let Y2 = I and X2 = I − X1 Y1 .
• The optimal controller is NOT uniquely defined.
• Don’t forget to check invertibility of I − D22 DK

Conclusion
The H∞ -optimal controller is a dynamic system.

AK B K
• Transfer Function K̂(s) =
CK DK
Minimizes the effect of external input (w) on external output (z).
kzkL2 ≤ kS(P, K)kH∞ kwkL2
• Minimum Energy Gain

H2 -optimal control
Motivation
H2 -optimal control minimizes the H2 -norm of the transfer function.

• The H2 -norm has no direct interpretation.
Z ∞
1
kGk2H2 = Trace(Ĝ(ıω)∗ Ĝ(ıω))dω
2π −∞
Motivation: Assume external input is Gaussian noise with signal variance Sw

Z ∞
1
E[w(t)2 ] = Trace(Ŝw (ıω))dω
2π −∞
Theorem 46.
For an LTI system P , if w is noise with spectral density Ŝw (ıω) and z = P w,
then z is noise with density
Ŝz (ıω) = P̂ (ıω)Ŝ(ıω)P̂ (ıω)∗

H2 -optimal control
Motivation
Then the output z = P w has signal variance (Power)

Z ∞
1
E[z(t)2 ] = Trace(Ĝ(ıω)∗ S(ıω)Ĝ(ıω))dω
2π −∞
≤ kSkH∞ kGk2H2
If the input signal is white noise, then Ŝ(ıω) = I and
E[z(t)2 ] = kGk2H2

H2 -optimal control
Colored Noise
Now suppose the noise is colored with variance Ŝw (ıω). Now define Ĥ as
Ĥ(ıω)Ĥ(ıω)∗ = Ŝw (ıω) and the filtered system.

P̂11 (s)Ĥ(s) P̂12 (s)
P̂s (s) =
P̂21 (s)Ĥ(s) P̂22 (s)
Now, applying feedback to the filtered plant, we get

S(Ps , K)(s) = P11 H + P12 (I − KP22 )−1 KP21 H = S(P, K)H
Now the spectral density, Ŝz of the output of the true plant using colored noise
equals the output of the artificial plant under white noise. i.e.
Sz (s) = S(P, K)(s)Ŝw (s)S(P, K)(s)∗
= S(P, K)(s)Ĥ(s)Ĥ(s)∗ S(P, K)(s)∗ = S(Ps , K)(s)Ŝ(Ps , K)(s)∗
Thus if K minimizes the H2 -norm of the filtered plant (kŜ(Ps , K)k2H2 ), it will
minimize the variance of the true plant under the influence of colored noise with
density Ŝw .

H2 -optimal control
Theorem 47.
Suppose P̂ (s) = C(sI − A)−1 B. Then the following are equivalent.
1. A is Hurwitz and kP̂ kH2 < γ.
2. There exists some X > 0 such that
trace CXC T < γ 2

AX + XAT + BB T < 0

H2 -optimal control
Proof.
Suppose A is Hurwitz and kP̂ kH2 < γ. Then the Controllability Grammian is
defined as Z ∞
T
Xc = eAt BB T eA dt
0
Now recall the Laplace transform
Z ∞

Λe At
(s) = eAt e−ts dt
Z0 ∞
= e−(sI−A)t dt
0
t=−∞

= −(sI − A) −1 −(sI−A)t
e dt
t=0
= (sI − A)−1

Hence ΛCeAt B (s) = C(sI − A)−1 B.

H2 -optimal control
Proof.

ΛCeAt B (s) = C(sI − A)−1 B implies
kP̂ k2H2 = kC(sI − A)−1 Bk2H2

Z ∞
1
= Trace((C(ıωI − A)−1 B)∗ (C(ıωI − A)−1 B))dω
2π 0
Z ∞
1
= Trace((C(ıωI − A)−1 B)(C(ıωI − A)−1 B)∗ )dω
2π 0
Z ∞
∗
= Trace CeAt BB ∗ eA t C ∗ dt
−∞
= TraceCXc C T
Thus Xc ≥ 0 and TraceCXc C T = kP̂ k2H2 < γ 2 .

H2 -optimal control
Proof.
Likewise TraceB T Xo B = kP̂ k2H2 where Xo is the observability Grammian. To
show that we can take strict the inequality X > 0, we simply let
Z ∞
T
X= eAt BB T + I eA dt
0
for sufficiently small > 0. Furthermore, we already know the controllability

grammian Xc and thus X satisfies the Lyapunov inequality.
AT X + X A + BB T < 0
These steps can be reversed to obtain necessity.

H2 -optimal control
Full-State Feedback
Lets consider the full-state feedback problem

 
A B1 B2
Ĝ(s) =  C1 0 D12 
I 0 0
• D12 is the weight on control effort.

• D11 = 0 is neglected as the feed-through term.
• C2 = I as this is state-feedback.

0 0
K̂(s) =
0 K

H2 -optimal control
Full-State Feedback
Theorem 48.
1. kS(K, P )kH2 < γ.
2. K = ZX −1 for some Z and X > 0 where

X AT
A B2 + X ZT + B1 B1T < 0
Z B2T

Trace C1 X + D12 Z X −1 C1 X + D12 Z < γ 2
However, this is nonlinear, so we need to reformulate using the Schur

Complement.

H2 -optimal control
Full-State Feedback
Theorem 49.
1. kS(K, P )kH2 < γ.
2. K = ZX −1 for some Z and X > 0 where

X AT
A B2 + X Z T
+ B1 B1T < 0
Z B2T

X (C1 X + D12 Z)T
>0
C1 X + D12 Z W
TraceW < γ 2
Thus we can solve the H2 -optimal static full-state feedback problem.

H2 -optimal control
Applying the Schur Complement gives the alternative formulation convenient for
control.
Theorem 50.
Suppose P̂ (s) = C(sI − A)−1 B. Then the following are equivalent.
1. A is Hurwitz and kP̂ kH2 < γ.
2. There exists some X, Z > 0 such that
T
A X + XA XB X CT
< 0, > 0, TraceZ < γ 2
BT X −γI C Z

H2 -optimal control
Relationship to LQR
The LQR Problem:

• Full-State Feedback
• Choose K to minimize the cost function
Z ∞
x(t)T Qx(t) + u(t)T Ru(t)dt
0
subject to dynamic constraints
ẋ(t) = Ax(t) + Bu(t)

u(t) = Kx(t), x(0) = x0

H2 -optimal control
Relationship to LQR
To solve the LQR problem using H2 optimal state-feedback control, let

1
Q2
• C1 =
0

0
• D12 = 1
R2
• B2 = B and B1 = I.
So that
 
A + BK I
A + B2 K B1
= 
1
S(P̂ , K̂) = Q2
C1 + D12 K D11 1 0
R K
2
And solve the H2 full-state feedback problem. Then if

ẋ(t) = ACL x(t) = (A + BK)x(t) = Ax(t) + Bu(t)
u(t) = Kx(t), x(0) = x0
Then x(t) = eACL t x0

H2 -optimal control
Relationship to LQR
If
ẋ(t) = ACL x(t) = (A + BK)x(t) = Ax(t) + Bu(t)

u(t) = Kx(t), x(0) = x0
then x(t) = eACL t x0 and

Z ∞ Z ∞
T
x(t)T Qx(t) + u(t)T Ru(t)dt = xT0 eACL t (Q + K T RK)eACL t x0 dt
0 0
Z ∞ 1 T 1
T Q 2 Q 2
= Trace xT0 eACL t 1 1 eACL t x0 dt
0 R2 K R2 K
Z ∞
T
= kx0 k2 Trace B1 eACL t (C1 + D12 K)T (C1 + D12 K)eACL t B1T dt
0
= kx0 k2 kS(K, P )k2H2
Thus LQR reduces to a special case of H2 static state-feedback.

H2 -optimal output feedback control
Theorem 51 (Lall).

AK B K
• There exists a K̂ = such that kS(K, P )kH2 < γ.
CK DK
• There exist X1 , Y1 , Z, An , Bn , Cn , Dn such that
 
AY1 +Y1 AT +B2 Cn +CnT B2T ∗T ∗T
 AT + An + [B2 Dn C2 ]T X1 A+AT X1 +Bn C2 +C2T BnT ∗T  < 0,
[B1 + B2 Dn D21 ]T [X1 B1 + Bn D21 ]T −I
 
Y1 I ∗T
 I X1 ∗T  > 0,
C1 Y1 + D12 Cn C1 + D12 Dn C2 Z
D11 + D12 Dn D21 = 0, trace(Z) < γ 2

H2 -optimal output feedback control
As before, the controller can be recovered as

−1 −1
AK2 BK2 X2 X1 B2 An B n X1 AY1 0 Y2T 0
= −
CK2 DK2 0 I C n D n 0 0 C2 Y1 I
for any full-rank X2 and Y2 such that
−1
X1 X2 Y1 Y2
=
X2T X3 Y2T Y3
To find the actual controller, we use the identities:
DK = (I + DK2 D22 )−1 DK2

BK = BK2 (I − D22 DK )
CK = (I − DK D22 )CK2
AK = AK2 − BK (I − D22 DK )−1 D22 CK

Robust Control
Before we finish, let us briefly touch on the use of LMIs in Robust Control.
p M q
Questions:
• Is S(∆, M ) stable for all ∆ ∈ ∆?
• Determine
sup kS(∆, M )kH∞ .
∆∈∆

Robust Control
Suppose we have the system M

M11 M12
M=
M21 M22
Definition 52.
We say the pair (M, ∆) is Robustly Stable if (I − M22 ∆) is invertible for all
∆ ∈ ∆.
Sl (M, ∆) = M11 + M12 ∆(I − M22 ∆)−1 M21

Robust Control
The structure of ∆ makes a lot of difference. e.g.

• Unstructured, Dynamic, norm-bounded:
∆ := {∆ ∈ L(L2 ) : k∆kH∞ < 1}
• Structured, Dynamic, norm-bounded:
∆ := {∆1 , ∆2 , · · · ∈ L(L2 ) : k∆i kH∞ < 1}
• Unstructured, Parametric, norm-bounded:
∆ := {∆ ∈ Rn×n : k∆k ≤ 1}
• Unstructured, Parametric, polytopic:

X X
∆ := {∆ ∈ Rn×n : ∆ = αi Hi , αi ≥ 0, αi ≤ 1}
i i

Robust Control
Let’s consider a simple question: Additive Uncertainty.
M11 = 0, M12 = M21 = I
Question: Is ẋ = A(t)x(t) stable if A(t) ∈ ∆ for all t ≥ 0.
Definition 53 (Quadratic Stability).

ẋ = A(t)x(t) is Quadratically Stable for A(t) ∈ ∆ if there exists some P > 0
such that
AT P + P A < 0 for all A ∈ ∆
Theorem 54.
If ẋ = A(t)x(t) is Quadratically Stable, then it is stable for A ∈ ∆.

Robust Control
We examine this problem for:

• Parametric, Polytopic Uncertainty:
X X
∆ := {∆ ∈ Rn×n : ∆ = αi Ai , αi ≥ 0, αi = 1}
i i

Parametric, Polytopic Uncertainty
For the polytopic case, we have the following result
Theorem 55 (Quadratic Stability).

Let X X
∆ := {∆ ∈ Rn×n : ∆ = αi Hi , αi ≥ 0, αi = 1}
i i
Then ẋ(t) = A(t)x(t) is quadratically stable for all A ∈ ∆ if and only if there
exists some P > 0 such that
ATi P + P Ai < 0 for i = 1, · · · ,
Thus quadratic stability of systems with polytopic uncertainty is equivalent to

an LMI.

Parametric, Norm-Bounded Uncertainty
A more complex uncertainty set is:
ẋ(t) = A0 x(t) + M p(t), p(t) = ∆(t)q(t),

q(t) = N x(t) + Qp(t), ∆∈∆
• Parametric, Norm-Bounded Uncertainty:
∆ := {∆ ∈ Rn×n : k∆k ≤ 1}

Quadratic Stability: There exists a P > 0 such that
P (A0 x(t)+M p)+(A0 x(t)+M p)T P < 0 for all p ∈ {p : p = ∆q, q = N x+Qp}
Theorem 56.
The system
ẋ(t) = A0 x(t) + M p(t), p(t) = ∆(t)q(t),

q(t) = N x(t) + Qp(t), ∆ ∈ ∆ := {∆ ∈ Rn×n : k∆k ≤ 1}
is quadratically stable if and only if there exists some P > 0 such that
T
x A P + PA PM x
<0
y MT P 0 y

x x x −N T N −N T Q x
for all ∈ : ≤0
y y y −QT N I − QT Q y

If.
If
T
x A P + PA PM x
<0
y MT P 0 y

x x x −N T N −N T Q x
for all ∈ : ≤ 0
y y y −QT N I − QT Q y
then
xT P (Ax + M y) + (Ax + M y)T P x < 0
for all x, y such that
kN x + Qyk2 ≤ kyk2
Therefore, since p = ∆q implies kpk ≤ kqk, we have quadratic stability.
The only if direction is similar.

Relationship to the S-Procedure
A Classical LMI
S-procedure to the rescue!

The S-procedure asks the question:
• Is z T F z ≥ 0 for all z ∈ {x : xT Gx ≥ 0}?
Corollary 57 (S-Procedure).
z T F z ≥ 0 for all z ∈ {x : xT Gx ≥ 0} if there exists a τ ≥ 0 such that
F − τ G 0.
The S-procedure is Necessary if {x : xT Gx > 0} =
6 ∅.

Theorem 58.
The system
ẋ(t) = Ax(t) + M p(t), p(t) = ∆(t)q(t),

q(t) = N x(t) + Qp(t), ∆ ∈ ∆ := {∆ ∈ Rn×n : k∆k ≤ 1}
is quadratically stable if and only if there exists some µ ≥ 0 and P > 0 such that

AP + P AT P N T MMT M QT
+µ < 0}
NP 0 QM T QQT − I
These approaches can be readily extended to controller synthesis.

Quadratic Stability
Consider Quadratic Stability in Discrete-Time: xk+1 = Sl (M, ∆)xk .
Definition 59.
(Sl , ∆) is QS if
Sl (M, ∆)T P Sl (M, ∆) − P < 0 for all ∆ ∈ ∆
Theorem 60 (Packard and Doyle).

Let M ∈ R(n+m)×(n+m) be given with ρ(M11 ) ≤ 1 and σ(M22 ) < 1. Then the
following are equivalent.
1. The pair (M, ∆ = Rm×m ) is quadratically stable.
2. The pair (M, ∆ = Cm×m ) is quadratically stable.
3. The pair (M, ∆ = Cm×m ) is robustly stable.

The Structured Singular Value
For the case of structured parametric uncertainty, we define the structured
singular value.
∆ = {∆ = diag(δ1 In1 , · · · , δs Ins , ∆s+1 , · · · , ∆s+f ) : δi ∈ F, ∆ ∈ Fnk ×nk }
• δ and ∆ represent unknown parameters.

• s is the number of scalar parameters.
• f is the number of matrix parameters.
Definition 61.
Given system M ∈ L(L2 ) and set ∆ as above, we define the Structured
Singular Value of (M, ∆) as
1
µ(M, ∆) =
inf ∆∈∆ k∆k
I−M22 ∆ is singular

The Structured Singular Value
Theorem 62.
Let
∆n = {∆ ∈ ∆, k∆k ≤ µ(M, ∆)}.
Then the pair (M, ∆n ) is robustly stable.

Conclusion
LMIs are a versatile tool for

• Optimal H∞ Control
• Optimal H2 Control (LQR/LQG)
• Robust Control
Next Lecture, we expand the use of LMIs exponentially
1. Nonlinear Systems Theory
2. Sum-of-Squares Nonlinear Stability Analysis
Time permitting, we will explore other applications
1. Stability and Control of Time-Delay Systems
2. Stability and Control of PDE systems.

L01 Mini PDF

Încărcat de

Informații document

Titlu original

Drepturi de autor

Formate disponibile

Partajați acest document

Partajați sau inserați document

Opțiuni de partajare

Vi se pare util acest document?

Este necorespunzător acest conținut?

Drepturi de autor:

Formate disponibile

L01 Mini PDF

Încărcat de

Drepturi de autor:

Formate disponibile

Introduction to Optimal Control via LMIs

Lecture 01: Optimal Control via LMIs

Powerful New Tools

Many old problems have been solved

Many questions are still unresolved

Question: What is meant by a “solution”?

M. Peet Lecture 01: 1 / 135

M. Peet Lecture 01: 2 / 135

L2 [0, ∞) inherits the norm

M. Peet Lecture 01: 3 / 135

• Satisfies the properties of a norm

if this limit exists.

M. Peet Lecture 01: 5 / 135

lim û(σ + ıω) = û(ıω)

I Which means continuous on the imaginary axis

I Which means integrable on every vertical line.

M. Peet Lecture 01: 6 / 135

Theorem 6 (Maximum Modulus).

Hence if û satisfies 1) and 2), then

We equip H2 with a norm and inner product

M. Peet Lecture 01: 7 / 135

• Shows that H2 is exactly the image of Λ on L2 [0, ∞)

where σ can be any real number.

M. Peet Lecture 01: 8 / 135

hΛu, ΛyiH2 = hu, yiL2

M. Peet Lecture 01: 9 / 135

• A Banach Space with norm

kĜkH∞ = ess sup σ̄(Ĝ(ıω))

M. Peet Lecture 01: 10 / 135

Every element of H∞ defines a multiplication operator.

(MĜ û)(s) = Ĝ(s)û(s)

Functions vs. Operators

M. Peet Lecture 01: 11 / 135

For any analytic functions, û and Ĝ, the function

H∞ is the space of transfer functions for linear time-invariant systems.

M. Peet Lecture 01: 12 / 135

From Paley-Wiener, if G = Λ−1 MĜ Λ

kGkL(L2 ) = kMĜ kL(H2 ) = kĜkH∞

The Gain of the system G can be calculated as kĜkH∞

Conclusion: H∞ provides a complete parametrization of the space of causal

M. Peet Lecture 01: 13 / 135

We define the following rational subspaces.

Note that RH2 and RH∞ are not complete(Banach) spaces.

M. Peet Lecture 01: 14 / 135

M. Peet Lecture 01: 15 / 135

For state-space system, (A, B, C, D),

z regulated outputs exogenous inputs w

We introduce the control framework by separating internal signals from external

outputs z are every output signal from the model.

z regulated outputs exogenous inputs w

I Not related to external input

s inputs w are all other inputs.

M. Peet Lecture 01: 18 / 135

The RegulatorExample: the regulator

M. Peet Lecture 01: 21 / 135

Command inputs and diagnostic outputs

M. Peet Lecture 01: 22 / 135

M. Peet Lecture 01: 23 / 135

z2 = u Wact Wsens nsensor = w3

Now we solve for z:

M. Peet Lecture 01: 26 / 135

on is called the (lower) star-product

M. Peet Lecture 01: 28 / 135

Because the rest is state-space,